# -*- coding: utf-8 -*- # Natural Language Toolkit: Language Model Unit Tests # # Copyright (C) 2001-2019 NLTK Project # Author: Ilia Kurenkov # URL: # For license information, see LICENSE.TXT from __future__ import division import math import sys import unittest from six import add_metaclass from nltk.lm import ( Vocabulary, MLE, Lidstone, Laplace, WittenBellInterpolated, KneserNeyInterpolated, ) from nltk.lm.preprocessing import padded_everygrams def _prepare_test_data(ngram_order): return ( Vocabulary(["a", "b", "c", "d", "z", "", ""], unk_cutoff=1), [ list(padded_everygrams(ngram_order, sent)) for sent in (list("abcd"), list("egadbe")) ], ) class ParametrizeTestsMeta(type): """Metaclass for generating parametrized tests.""" def __new__(cls, name, bases, dct): contexts = ( ("a",), ("c",), (u"",), ("b",), (u"",), ("d",), ("e",), ("r",), ("w",), ) for i, c in enumerate(contexts): dct["test_sumto1_{0}".format(i)] = cls.add_sum_to_1_test(c) scores = dct.get("score_tests", []) for i, (word, context, expected_score) in enumerate(scores): dct["test_score_{0}".format(i)] = cls.add_score_test( word, context, expected_score ) return super(ParametrizeTestsMeta, cls).__new__(cls, name, bases, dct) @classmethod def add_score_test(cls, word, context, expected_score): if sys.version_info > (3, 5): message = "word='{word}', context={context}" else: # Python 2 doesn't report the mismatched values if we pass a custom # message, so we have to report them manually. message = ( "{score} != {expected_score} within 4 places, " "word='{word}', context={context}" ) def test_method(self): score = self.model.score(word, context) self.assertAlmostEqual( score, expected_score, msg=message.format(**locals()), places=4 ) return test_method @classmethod def add_sum_to_1_test(cls, context): def test(self): s = sum(self.model.score(w, context) for w in self.model.vocab) self.assertAlmostEqual(s, 1.0, msg="The context is {}".format(context)) return test @add_metaclass(ParametrizeTestsMeta) class MleBigramTests(unittest.TestCase): """unit tests for MLENgramModel class""" score_tests = [ ("d", ["c"], 1), # Unseen ngrams should yield 0 ("d", ["e"], 0), # Unigrams should also be 0 ("z", None, 0), # N unigrams = 14 # count('a') = 2 ("a", None, 2.0 / 14), # count('y') = 3 ("y", None, 3.0 / 14), ] def setUp(self): vocab, training_text = _prepare_test_data(2) self.model = MLE(2, vocabulary=vocab) self.model.fit(training_text) def test_logscore_zero_score(self): # logscore of unseen ngrams should be -inf logscore = self.model.logscore("d", ["e"]) self.assertTrue(math.isinf(logscore)) def test_entropy_perplexity_seen(self): # ngrams seen during training trained = [ ("", "a"), ("a", "b"), ("b", ""), ("", "a"), ("a", "d"), ("d", ""), ] # Ngram = Log score # , a = -1 # a, b = -1 # b, UNK = -1 # UNK, a = -1.585 # a, d = -1 # d, = -1 # TOTAL logscores = -6.585 # - AVG logscores = 1.0975 H = 1.0975 perplexity = 2.1398 self.assertAlmostEqual(H, self.model.entropy(trained), places=4) self.assertAlmostEqual(perplexity, self.model.perplexity(trained), places=4) def test_entropy_perplexity_unseen(self): # In MLE, even one unseen ngram should make entropy and perplexity infinite untrained = [("", "a"), ("a", "c"), ("c", "d"), ("d", "")] self.assertTrue(math.isinf(self.model.entropy(untrained))) self.assertTrue(math.isinf(self.model.perplexity(untrained))) def test_entropy_perplexity_unigrams(self): # word = score, log score # = 0.1429, -2.8074 # a = 0.1429, -2.8074 # c = 0.0714, -3.8073 # UNK = 0.2143, -2.2224 # d = 0.1429, -2.8074 # c = 0.0714, -3.8073 # = 0.1429, -2.8074 # TOTAL logscores = -21.6243 # - AVG logscores = 3.0095 H = 3.0095 perplexity = 8.0529 text = [("",), ("a",), ("c",), ("-",), ("d",), ("c",), ("",)] self.assertAlmostEqual(H, self.model.entropy(text), places=4) self.assertAlmostEqual(perplexity, self.model.perplexity(text), places=4) @add_metaclass(ParametrizeTestsMeta) class MleTrigramTests(unittest.TestCase): """MLE trigram model tests""" score_tests = [ # count(d | b, c) = 1 # count(b, c) = 1 ("d", ("b", "c"), 1), # count(d | c) = 1 # count(c) = 1 ("d", ["c"], 1), # total number of tokens is 18, of which "a" occured 2 times ("a", None, 2.0 / 18), # in vocabulary but unseen ("z", None, 0), # out of vocabulary should use "UNK" score ("y", None, 3.0 / 18), ] def setUp(self): vocab, training_text = _prepare_test_data(3) self.model = MLE(3, vocabulary=vocab) self.model.fit(training_text) @add_metaclass(ParametrizeTestsMeta) class LidstoneBigramTests(unittest.TestCase): """unit tests for Lidstone class""" score_tests = [ # count(d | c) = 1 # *count(d | c) = 1.1 # Count(w | c for w in vocab) = 1 # *Count(w | c for w in vocab) = 1.8 ("d", ["c"], 1.1 / 1.8), # Total unigrams: 14 # Vocab size: 8 # Denominator: 14 + 0.8 = 14.8 # count("a") = 2 # *count("a") = 2.1 ("a", None, 2.1 / 14.8), # in vocabulary but unseen # count("z") = 0 # *count("z") = 0.1 ("z", None, 0.1 / 14.8), # out of vocabulary should use "UNK" score # count("") = 3 # *count("") = 3.1 ("y", None, 3.1 / 14.8), ] def setUp(self): vocab, training_text = _prepare_test_data(2) self.model = Lidstone(0.1, 2, vocabulary=vocab) self.model.fit(training_text) def test_gamma(self): self.assertEqual(0.1, self.model.gamma) def test_entropy_perplexity(self): text = [ ("", "a"), ("a", "c"), ("c", ""), ("", "d"), ("d", "c"), ("c", ""), ] # Unlike MLE this should be able to handle completely novel ngrams # Ngram = score, log score # , a = 0.3929, -1.3479 # a, c = 0.0357, -4.8074 # c, UNK = 0.0(5), -4.1699 # UNK, d = 0.0263, -5.2479 # d, c = 0.0357, -4.8074 # c, = 0.0(5), -4.1699 # TOTAL logscore: −24.5504 # - AVG logscore: 4.0917 H = 4.0917 perplexity = 17.0504 self.assertAlmostEqual(H, self.model.entropy(text), places=4) self.assertAlmostEqual(perplexity, self.model.perplexity(text), places=4) @add_metaclass(ParametrizeTestsMeta) class LidstoneTrigramTests(unittest.TestCase): score_tests = [ # Logic behind this is the same as for bigram model ("d", ["c"], 1.1 / 1.8), # if we choose a word that hasn't appeared after (b, c) ("e", ["c"], 0.1 / 1.8), # Trigram score now ("d", ["b", "c"], 1.1 / 1.8), ("e", ["b", "c"], 0.1 / 1.8), ] def setUp(self): vocab, training_text = _prepare_test_data(3) self.model = Lidstone(0.1, 3, vocabulary=vocab) self.model.fit(training_text) @add_metaclass(ParametrizeTestsMeta) class LaplaceBigramTests(unittest.TestCase): """unit tests for Laplace class""" score_tests = [ # basic sanity-check: # count(d | c) = 1 # *count(d | c) = 2 # Count(w | c for w in vocab) = 1 # *Count(w | c for w in vocab) = 9 ("d", ["c"], 2.0 / 9), # Total unigrams: 14 # Vocab size: 8 # Denominator: 14 + 8 = 22 # count("a") = 2 # *count("a") = 3 ("a", None, 3.0 / 22), # in vocabulary but unseen # count("z") = 0 # *count("z") = 1 ("z", None, 1.0 / 22), # out of vocabulary should use "UNK" score # count("") = 3 # *count("") = 4 ("y", None, 4.0 / 22), ] def setUp(self): vocab, training_text = _prepare_test_data(2) self.model = Laplace(2, vocabulary=vocab) self.model.fit(training_text) def test_gamma(self): # Make sure the gamma is set to 1 self.assertEqual(1, self.model.gamma) def test_entropy_perplexity(self): text = [ ("", "a"), ("a", "c"), ("c", ""), ("", "d"), ("d", "c"), ("c", ""), ] # Unlike MLE this should be able to handle completely novel ngrams # Ngram = score, log score # , a = 0.2, -2.3219 # a, c = 0.1, -3.3219 # c, UNK = 0.(1), -3.1699 # UNK, d = 0.(09), 3.4594 # d, c = 0.1 -3.3219 # c, = 0.(1), -3.1699 # Total logscores: −18.7651 # - AVG logscores: 3.1275 H = 3.1275 perplexity = 8.7393 self.assertAlmostEqual(H, self.model.entropy(text), places=4) self.assertAlmostEqual(perplexity, self.model.perplexity(text), places=4) @add_metaclass(ParametrizeTestsMeta) class WittenBellInterpolatedTrigramTests(unittest.TestCase): def setUp(self): vocab, training_text = _prepare_test_data(3) self.model = WittenBellInterpolated(3, vocabulary=vocab) self.model.fit(training_text) score_tests = [ # For unigram scores by default revert to MLE # Total unigrams: 18 # count('c'): 1 ("c", None, 1.0 / 18), # in vocabulary but unseen # count("z") = 0 ("z", None, 0.0 / 18), # out of vocabulary should use "UNK" score # count("") = 3 ("y", None, 3.0 / 18), # gamma(['b']) = 0.1111 # mle.score('c', ['b']) = 0.5 # (1 - gamma) * mle + gamma * mle('c') ~= 0.45 + .3 / 18 ("c", ["b"], (1 - 0.1111) * 0.5 + 0.1111 * 1 / 18), # building on that, let's try 'a b c' as the trigram # gamma(['a', 'b']) = 0.0667 # mle("c", ["a", "b"]) = 1 ("c", ["a", "b"], (1 - 0.0667) + 0.0667 * ((1 - 0.1111) * 0.5 + 0.1111 / 18)), ] @add_metaclass(ParametrizeTestsMeta) class KneserNeyInterpolatedTrigramTests(unittest.TestCase): def setUp(self): vocab, training_text = _prepare_test_data(3) self.model = KneserNeyInterpolated(3, vocabulary=vocab) self.model.fit(training_text) score_tests = [ # For unigram scores revert to uniform # Vocab size: 8 # count('c'): 1 ("c", None, 1.0 / 8), # in vocabulary but unseen, still uses uniform ("z", None, 1 / 8), # out of vocabulary should use "UNK" score, i.e. again uniform ("y", None, 1.0 / 8), # alpha = count('bc') - discount = 1 - 0.1 = 0.9 # gamma(['b']) = discount * number of unique words that follow ['b'] = 0.1 * 2 # normalizer = total number of bigrams with this context = 2 # the final should be: (alpha + gamma * unigram_score("c")) ("c", ["b"], (0.9 + 0.2 * (1 / 8)) / 2), # building on that, let's try 'a b c' as the trigram # alpha = count('abc') - discount = 1 - 0.1 = 0.9 # gamma(['a', 'b']) = 0.1 * 1 # normalizer = total number of trigrams with prefix "ab" = 1 => we can ignore it! ("c", ["a", "b"], 0.9 + 0.1 * ((0.9 + 0.2 * (1 / 8)) / 2)), ] class NgramModelTextGenerationTests(unittest.TestCase): """Using MLE estimator, generate some text.""" def setUp(self): vocab, training_text = _prepare_test_data(3) self.model = MLE(3, vocabulary=vocab) self.model.fit(training_text) def test_generate_one_no_context(self): self.assertEqual(self.model.generate(random_seed=3), "") def test_generate_one_limiting_context(self): # We don't need random_seed for contexts with only one continuation self.assertEqual(self.model.generate(text_seed=["c"]), "d") self.assertEqual(self.model.generate(text_seed=["b", "c"]), "d") self.assertEqual(self.model.generate(text_seed=["a", "c"]), "d") def test_generate_one_varied_context(self): # When context doesn't limit our options enough, seed the random choice self.assertEqual( self.model.generate(text_seed=("a", ""), random_seed=2), "a" ) def test_generate_cycle(self): # Add a cycle to the model: bd -> b, db -> d more_training_text = [list(padded_everygrams(self.model.order, list("bdbdbd")))] self.model.fit(more_training_text) # Test that we can escape the cycle self.assertEqual( self.model.generate(7, text_seed=("b", "d"), random_seed=5), ["b", "d", "b", "d", "b", "d", ""], ) def test_generate_with_text_seed(self): self.assertEqual( self.model.generate(5, text_seed=("", "e"), random_seed=3), ["", "a", "d", "b", ""], ) def test_generate_oov_text_seed(self): self.assertEqual( self.model.generate(text_seed=("aliens",), random_seed=3), self.model.generate(text_seed=("",), random_seed=3), ) def test_generate_None_text_seed(self): # should crash with type error when we try to look it up in vocabulary with self.assertRaises(TypeError): self.model.generate(text_seed=(None,)) # This will work self.assertEqual( self.model.generate(text_seed=None, random_seed=3), self.model.generate(random_seed=3), )